I am playing with a dataset and a logistic regression program that was developed by my predecessors where I work. I am convinced that the model they are using is hugely overfit ..... there's considerable evidence, like huge confidence intervals around odds ratios, in addition to radical violation of rules of thumb about variable/subject ratios.
But I noticed soemthing else: In the overfit model, the Wald test of the global null hypothesis is not significant, while the LR and Score tests are highly significant.
Is this a general phenomenon?
Peter L. Flom, PhD
www DOT peterflom DOT com